[mikasenghaas]

Summary

• generate() returns completion_logprobs as a float32 numpy array, not a list of Python floats.
• prime-rl's packed form (a base64 float32 buffer) now decodes without a copy. Before, generate() widened it to float64 and built a list. verifiers then narrowed that list back into a float32 array.
• The JSON and compact forms also return float32 arrays. vLLM samples in float32, so this narrowing is lossless.

Part of a set with PrimeIntellect-ai/verifiers#2961 and PrimeIntellect-ai/prime-rl#4226. verifiers accepts either a list or an array, so the PRs can merge in any order.

Breaking

• generate()["completion_logprobs"] is a np.ndarray (float32), not a list[float]. Code that compares it to a list with ==, or tests its truthiness, must use .tolist(), len() or pytest.approx. Indexing, len() and iteration work as before.

Tests

• tests/test_client.py and tests/replay.py compare the array through .tolist() or pytest.approx, and check the float32 dtype.
• tests/test_client.py and tests/test_incremental.py pass locally.
• Full suite: 10 failures, all test_multimodal.py Inkling processor-parity cases. They render prompts and don't call generate().

🤖 Generated with Claude Code

---

Note

Medium Risk
Breaking API change for generate()["completion_logprobs"] (list → ndarray); downstream code that assumed list equality or truthiness may fail until updated.

Overview
generate() now returns completion_logprobs as a float32 numpy array instead of a list[float], avoiding extra widen-to-float64 / list materialization on the hot path (especially for prime-rl’s base64 packed float32 buffer, which can decode without a copy).

The logprob parsers (_parse_packed_completion_logprobs, _parse_compact_logprobs, _parse_completion_logprobs) and compact sampling output now build or return np.ndarray dtype float32; generate()’s docstring documents the new shape. Breaking: callers that compared the field to a plain list with == need .tolist(), pytest.approx, or element-wise checks; indexing, len(), and iteration still work.

Tests in tests/test_client.py and tests/replay.py assert dtype and compare via .tolist() / pytest.approx.

^{Reviewed by Cursor Bugbot for commit faf9e2e. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Return completion logprobs from renderers as float32 NumPy arrays

• _parse_packed_completion_logprobs, _parse_compact_logprobs, and _parse_completion_logprobs in renderers/client.py now return float32 NumPy arrays instead of Python float lists
• Avoids a copy in the packed path when the source dtype already permits it; validation of token counts, finiteness, and the vLLM sentinel is unchanged
• Updates tests in tests/test_client.py and tests/replay.py to compare array results via pytest.approx or list conversion, and adds dtype checks
• Behavioral Change: completion_logprobs is now a float32 NumPy array with one entry per completion token, so consumers expecting Python floats lose precision and list semantics

^{Macroscope summarized faf9e2e.}

[macroscopeapp]

Approvability

Verdict: Needs Human Approval

Macroscope's review found this PR requires human approval — The PR improves allocation efficiency by returning float32 NumPy arrays, but changes the existing public completion_logprobs contract from list[float] to np.ndarray. Downstream callers may require updates for list-specific semantics, making compatibility review important.

Notes:

• Macroscope's correctness review did not run, so approvability was decided on eligibility alone.

^{You can add or adjust custom eligibility rules. Learn more.}